DeepMSP

Introduction

Official code associated with the following publication:

Ari Tchetchenian, Leo Zekelman, Yuqian Chen, Jarrett Rushmore, Fan Zhang, Edward H. Yeterian, Nikos Makris, Yogesh Rathi, Erik Meijering, Yang Song, Lauren J. O'Donnell. (2024). "Deep multimodal saliency parcellation of cerebellar pathways: linking microstructure and individual function through explainable multitask learning", Human Brain Mapping. https://doi.org/10.1002/hbm.70008

Setup

Step 1: Download Required Files

Download the following files required for the project. See 'File Structure Details/Examples' for more details:

• clusters.csv: A list of fiber cluster IDs, e.g. "left_hemisphere.cluster_12345".
• HCP_n1065_allDWI_fiber_clusters.csv: A set of DWI measurements for each cluster of each subject.
• S1200_demographics_Behavioral.csv: A set of individual functional performance measures for each subject

Step 2: Create Dataset

Run the following command to create the dataset needed for training:

python create_dataset.py

Step 3: Generate CSV Files

To generate the .csv files needed for the project, execute:

python generate_csvs.py

Usage

Training the Model

To train the model, run train_multi.py with your desired arguments, for example:

python train_multi.py --batch_size 10 --epochs 50 --learning_rate 1e-4 --input_channels 1940 --results_dir ./results --dataset_dir new_dataset --dropout 10 --save_name cerebellum_optimised_transformer

Performing Clustering

After training, perform parcellation by running explain_multi.py with your desired arguments, for example:

python explain_multi.py --save_name clustering_results --results_name cerebellum_optimised_transformer --model transformer --bilateral

File Structure Details/Examples

clusters.csv

A simple headerless .csv file where each row contains a single cluster ID string.

left_hemisphere.cluster_12345
right_hemisphere.cluster_12345
commissural.cluster_54321
...

HCP_n1065_allDWI_fiber_clusters.csv

Each row contains a subject ID followed by a variety of DWI measures for each cluster ID of that subject. The create_dataset.py script assumes this .csv file contains columns for the Min/Max/Median/Mean/Variance for: FA1, FA2, Num_Fibers, Num_Points; Min/Max/Median/Mean for: correct_trace1, correct_trace2; and additionally: Num_Fibers, Num_Points columns.

subjectkey	left_hemisphere.cluster_12345.Num_Fibers	left_hemisphere.cluster_12345.FA1.Mean	...	right_hemisphere.cluster_12345.NumFibers	...
555555	304	0.34	...	403	...
333333	201	0.55	...	204	...
444444	111	0.32	...	112	...
...	...	...	...	...	...

S1200_demographics_Behavioral.csv

This file is a behavioral data file available in the "WU-Minn HCP Data - 1200 Subjects" dataset.

Each row contains a set of functional/behavioral measures for the corresponding subject. If you are using functional/behavioral measures that are not used by the HCP-YA dataset, you may need to alter the names of the headers in the generate_csvs.py file to match your dataset's headers.

Subject	PicVocab_AgeAdj	ReadEng_AgeAdj	...
555555	102	85	...
333333	99	94	...
444444	105	83	...
...	...	...	...